Super-Resolution Land Cover Mapping Using Multiscale Self-Similarity Redundancy

Super-resolution mapping (SRM) is a method to estimate a fine-resolution land cover map from coarse-resolution fraction images. SRM is an ill-posed problem and regularization terms are always needed to be introduced to well-pose the solution. The regularization term based on the maximal spatial dependence has been widely used in SRM; however, it is often too simple to provide detailed land cover pattern information. In this paper, a novel SRM algorithm, which adopts a new regularization term that is generated using the multiscale self-similarity redundancy feature in fraction images, is proposed. Based on the multiscale self-similarity redundancy, the proposed SRM algorithm magnifies the input coarse-resolution fraction images to the images with the same spatial resolution as the final fine-resolution map to be estimated. A coarse-to-fine strategy is applied to insure the stability of the magnifying process. During the magnifying process, support vector regression is used to learn the relationship between image patches in fraction images and the down-sampled images, which is applied to estimate the magnified fraction images. The new regularization term is then constructed based on the final magnified fraction images and used to provide additional information for SRM. The proposed SRM algorithm was compared with popular SRM algorithms using both synthetic and real images. Experimental results show that the multiscale self-similarity redundancy feature widely exists in fraction images, and provides valuable land cover information for SRM. Resultant fine-resolution land cover maps generated by the proposed SRM algorithm have higher accuracy than other algorithms.